DragMesh-2: Physically Plausible Dexterous Hand-Object Interaction with Articulated Objects

This is the official repository for the paper:

DragMesh-2: Physically Plausible Dexterous Hand-Object Interaction with Articulated Objects

Tianshan Zhang*, Yijia Duan*, Yanjun Li*, Zeyu Zhang*^†, and Hao Tang^‡

School of Computer Science, Peking University

*Equal contribution. ^†Project lead. ^‡Corresponding author: bjdxtanghao@gmail.com.

Paper | Website | Models | Data

Note

The simulation demo and real-robot demo are intentionally left as placeholders in this release. They will be added after the corresponding experiments are finalized.

teaser.mp4

Citation

If you find our code or paper useful, please consider citing:

@article{zhang2026dragmesh2,
title={DragMesh-2: Physically Plausible Dexterous Hand-Object Interaction with Articulated Objects},
author={Zhang, Tianshan and Duan, Yijia and Li, Yanjun and Zhang, Zeyu and Tang, Hao},
journal={arXiv preprint arXiv:2606.15133},
year={2026}
}

Intro

DragMesh-2 is a contact-driven framework for dexterous hand interaction with articulated objects. It extends articulated interaction from object-centric generation to hand-driven hand-object interaction, where a target part can move only through physical contact between the dexterous hand and the handle.

The repository contains:

• an Isaac Gym articulated-object interaction environment built around a 51-DoF SMPL-X right hand;
• PPO training code with a separated PICA method module and GLA backbone;
• deterministic and stochastic evaluation tools for damping and contact-load robustness;
• trajectory-tracking baseline scripts.

News

2026/06/01: Initial public-code cleanup for DragMesh-2.

TODO List

• Release core simulation and PPO/PICA code.
• Add cleaned training, evaluation, and baseline entrypoints.
• Add release README and setup instructions.
• Release prepared trajectory/data package.
• Release pretrained checkpoints.
• Add simulation demo.
• Add real-robot demo.

Quick Start

Environment Setup

The code is designed for Isaac Gym preview environments. Our internal runs use Python 3.8/3.9, CUDA-capable PyTorch, and Isaac Gym.

Install Python dependencies:

pip install -r requirements.txt

Install external packages that are not distributed through this repository:

# Isaac Gym# Download Isaac Gym from NVIDIA, then add it to PYTHONPATH.export ISAACGYM_ROOT=/path/to/isaacgym
export PYTHONPATH="${ISAACGYM_ROOT}/python:${PYTHONPATH}"# GLA encoder dependency used by PICA checkpoints.# Install flash-linear-attention or keep a local checkout at ./flash-linear-attention.export FLA_ROOT=/path/to/flash-linear-attention
export PYTHONPATH="${FLA_ROOT}:${PYTHONPATH}"

If you use a conda environment, also expose its runtime libraries:

export LD_LIBRARY_PATH="${CONDA_PREFIX}/lib:${LD_LIBRARY_PATH}"

Data Preparation

Prepare GAPartNet assets in this layout:

assets/
`-- gapartnet_example/
`-- 45661/
|-- mobility_annotation_gapartnet.urdf
|-- mobility_v2.json
`-- link_annotation_gapartnet.json

The default config expects:

hand.hand_asset_root: hands/floatingasset.asset_root: assetsasset.arti_obj_root: gapartnet_example

The repository already includes smplx_right_hand_floating.urdf. If you need to rebuild it from a base SMPL-X right-hand URDF:

python dataset/geometric/build_hand_urdf.py \
--input hands/smplx_variants/hands_only/smplx_right_hand.urdf \
--output hands/floating/smplx_right_hand_floating.urdf

Generate Contact Trajectories

Edit configs/env/hand_config.yaml to select the GAPartNet object IDs, then run:

python dataset/geometric/run_hand_drag.py --config configs/env/hand_config.yaml --headless

Generated trajectories are written under:

output/hand_drag/<object_id>/trajectory.json

For batch training/evaluation, create a manifest at data/manifest.csv:

sample_id,object_id,trajectory,enabled45661,45661,output/hand_drag/45661/trajectory.json,1

You can also pass --trajectory directly to all training and evaluation commands.

Training

Train the main PICA configuration:

OBJECT_ID=45661 \
TRAJECTORY=output/hand_drag/45661/trajectory.json \
RUN_NAME=dragmesh2_45661_pica \
bash scripts/train/train_pica.sh

Equivalent direct command:

python ppo/train.py \
--train_config configs/train/pica/train_config_gla_pica_drand12_aux_v2c.yaml \
--object_id 45661 \
--trajectory output/hand_drag/45661/trajectory.json \
--num_envs 64 \
--max_epochs 150 \
--experiment_name dragmesh2_45661_pica

Checkpoints are saved under:

runs/<experiment_name>/nn/

Evaluation

Run deterministic and stochastic damping evaluation:

OBJECT_ID=45661 \
TRAJECTORY=output/hand_drag/45661/trajectory.json \
CHECKPOINT=runs/dragmesh2_45661_pica/nn \
OUT_ROOT=output/eval/dragmesh2_45661_pica \
bash scripts/eval/eval_det_stoch_damping.sh

Run a single deterministic evaluation:

python scripts/evaluate_ppo_baseline.py \
--object_id 45661 \
--trajectory output/hand_drag/45661/trajectory.json \
--checkpoint runs/dragmesh2_45661_pica/nn \
--checkpoint-kind best \
--episodes 20 \
--summary_json output/eval/45661_summary.json

Baselines

Run the non-learned trajectory-tracking baseline:

python scripts/track_trajectory_baseline.py \
--object_id 45661 \
--trajectory output/hand_drag/45661/trajectory.json \
--phase drag \
--summary_json output/baselines/45661_tracking.json

Batch baseline evaluation over data/manifest.csv:

python scripts/run_trajectory_baselines.py \
--manifest data/manifest.csv \
--phase drag

Repository Layout

.
|-- dataset/geometric/ # Geometry-only dataset generator
|-- hands/ # Dexterous hand assets
|-- configs/ # Environment and training YAML files
|-- backbones/ # GLA backbone plus MLP/GRU/Transformer notes
|-- pica/ # PICA agent extension and method notes
|-- ppo/ # PPO trainer, task, and rl-games environment wrapper
|-- scripts/
| |-- train/ # Standard training wrappers
| |-- eval/ # Standard evaluation wrappers
| `-- baselines/ # Baseline wrappers
|-- assets/ # External articulated assets live here
|-- data/ # Manifest templates and local data pointers
`-- demos/ # Simulation and real-robot demos, coming soon

Acknowledgement

We thank the authors of Isaac Gym, GAPartNet, SMPL-X, rl-games, and flash-linear-attention for their open-source code and datasets.